A growing set of open, self-contained prompts that make Claude Code work without you in the loop — improving itself after every session, and acting on its own when you mention it on a GitHub issue or PR. Paste one into Claude Code and it builds the capability into your repo as files you own and edit.
Claude that gets better on its own. When a working session ends, a Stop
hook reviews the work and commits what would make the next run faster — as skills in
your project's own .claude/skills/. No prompting, no curation from you.
How to use the project's libraries and conventions, indexed as a table of contents. Anchored by a self-healing project-overview.
A generic CLI that drives the real interface — UI, command, endpoint, API — so a result is proven the way its user sees it. Never a proxy.
A generic CLI over an external system — a DB, an issue tracker, an API. The one gap that may need you: authentication.
# paste into Claude Code in your repo — it writes the Stop hook + # review skill into .claude/, then tells you to /clear
@claude agent works without youClaude that works without you. Mention @claude … on a GitHub
issue and a session in Anthropic's cloud does the work and opens a PR — your laptop
off, on your Max subscription. Two pieces, set up once.
Write @claude … in an issue or comment. A tiny dispatcher Action you own checks permissions and fires a cloud routine — it does zero Claude work itself.
The routine clones the repo, implements the change, runs the tests, and opens a pull request — your laptop off, on your Max subscription.
From there Auto-fix watches the PR and handles review comments and CI on its own — no re-triggering needed.
# paste into Claude Code in your repo — it writes the dispatcher Action into # .github/, then prints the one-time routine + secrets setup for you to do in the web UI
Real secrets in your routine, never in your repo. The cloud environment
holds one bootstrap token for your secrets manager; a SessionStart hook in
your repo pulls the rest and injects them into the session at runtime — so the agent
runs with live API keys and connection strings, and nothing sensitive is ever committed.
Your routine environment holds a single secrets-manager token as an env var. Every other secret stays in the manager — Doppler, Vault, Infisical, 1Password, AWS.
The environment's setup script installs the manager's CLI, so it's ready before any session starts.
A SessionStart hook you own runs the CLI and writes the secrets into the session's env file — available to every command, never to disk you commit.
# paste into Claude Code in your repo — it asks which secrets manager you use, # writes the SessionStart hook into .claude/, then prints the web-UI steps for the environment